Hamming Loss implementation in Keras

[Sudheer Kumar]

Is there an example of Hamming Loss implementation in Keras for multi-label classification? Also wanted to see if anyone can share their experience with hamming noise loss function for multi-label classification. 

Is it possible to use SciKit implementation of Hamming Noise loss function with Keras?

With Best Regards

Sudheer Kumar

[Lance N.]

I found this on stackoverflow, but I have not tried it.

def hamming_loss(y_true, y_pred):

        return K.mean(y_true*(1-y_pred)+(1-y_true)*y_pred)

[Sudheer Kumar]

Thank you, Lance. I found another implementation by Manish on Medium. I wish Keras would provide some of these functions in the upcoming releases. I used this function but I didn't see a huge lift compared BCE. 

# CREDITS

#https://blog.manash.me/multi-task-learning-in-keras-implementation-of-multi-task-classification-loss-f1d42da5c3f6

#https://blog.manash.me/@ManashMandal

import keras.backend as K

def hn_multilabel_loss(y_true, y_pred):

    # Avoid divide by 0

    y_pred = K.clip(y_pred, K.epsilon(), 1 - K.epsilon())

    # Multi-task loss

    return K.mean(K.sum(- y_true * K.log(y_pred) - (1 - y_true) * K.log(1 - y_pred), axis=1))

[ckm...@gmail.com]

Hi, what Manish has implemented is cross entropy loss over the entire data. Hamming loss is something which computes the difference between true labels and predicted labels. See here, https://scikit-learn.org/stable/modules/model_evaluation.html#hamming-loss

[ckm...@gmail.com]

Here is my imp. I assume that y_pred is coming from a sigmoid layer without thresholding at 0.5

def Hamming_loss(y_true, y_pred):

    tmp = K.abs(y_true-y_pred)

    return K.mean(K.cast(K.greater(tmp,0.5),dtype=float))

- show quoted text -